Migrate metric to Evaluate in Pytorch examples (#18369)

* Migrate metric to Evaluate in pytorch examples

* Remove unused imports

examples/pytorch/_tests_requirements.txt

@@ -23,3 +23,4 @@ torchvision
 jiwer
 librosa
 torch < 1.12
+evaluate

examples/pytorch/audio-classification/run_audio_classification.py

@@ -26,6 +26,7 @@
 import numpy as np
 from datasets import DatasetDict, load_dataset
 
+import evaluate
 import transformers
 from transformers import (
     AutoConfig,
@@ -315,7 +316,7 @@ def val_transforms(batch):
         id2label[str(i)] = label
 
     # Load the accuracy metric from the datasets package
-    metric = datasets.load_metric("accuracy")
+    metric = evaluate.load("accuracy")
 
     # Define our compute_metrics function. It takes an `EvalPrediction` object (a namedtuple with
     # `predictions` and `label_ids` fields) and has to return a dictionary string to float.

examples/pytorch/image-classification/run_image_classification.py

@@ -19,7 +19,6 @@
 from dataclasses import dataclass, field
 from typing import Optional
 
-import datasets
 import numpy as np
 import torch
 from datasets import load_dataset
@@ -34,6 +33,7 @@
     ToTensor,
 )
 
+import evaluate
 import transformers
 from transformers import (
     MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING,
@@ -252,7 +252,7 @@ def main():
         id2label[str(i)] = label
 
     # Load the accuracy metric from the datasets package
-    metric = datasets.load_metric("accuracy")
+    metric = evaluate.load("accuracy")
 
     # Define our compute_metrics function. It takes an `EvalPrediction` object (a namedtuple with a
     # predictions and label_ids field) and has to return a dictionary string to float.

examples/pytorch/image-classification/run_image_classification_no_trainer.py

@@ -22,7 +22,7 @@
 
 import datasets
 import torch
-from datasets import load_dataset, load_metric
+from datasets import load_dataset
 from torch.utils.data import DataLoader
 from torchvision.transforms import (
     CenterCrop,
@@ -35,6 +35,7 @@
 )
 from tqdm.auto import tqdm
 
+import evaluate
 import transformers
 from accelerate import Accelerator
 from accelerate.logging import get_logger
@@ -415,7 +416,7 @@ def collate_fn(examples):
             accelerator.init_trackers("image_classification_no_trainer", experiment_config)
 
     # Get the metric function
-    metric = load_metric("accuracy")
+    metric = evaluate.load("accuracy")
 
     # Train!
     total_batch_size = args.per_device_train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps

examples/pytorch/language-modeling/run_clm.py

@@ -30,8 +30,9 @@
 from typing import Optional
 
 import datasets
-from datasets import load_dataset, load_metric
+from datasets import load_dataset
 
+import evaluate
 import transformers
 from transformers import (
     CONFIG_MAPPING,
@@ -492,7 +493,7 @@ def preprocess_logits_for_metrics(logits, labels):
                 logits = logits[0]
             return logits.argmax(dim=-1)
 
-        metric = load_metric("accuracy")
+        metric = evaluate.load("accuracy")
 
         def compute_metrics(eval_preds):
             preds, labels = eval_preds

examples/pytorch/language-modeling/run_mlm.py

@@ -30,8 +30,9 @@
 from typing import Optional
 
 import datasets
-from datasets import load_dataset, load_metric
+from datasets import load_dataset
 
+import evaluate
 import transformers
 from transformers import (
     CONFIG_MAPPING,
@@ -515,7 +516,7 @@ def preprocess_logits_for_metrics(logits, labels):
                 logits = logits[0]
             return logits.argmax(dim=-1)
 
-        metric = load_metric("accuracy")
+        metric = evaluate.load("accuracy")
 
         def compute_metrics(eval_preds):
             preds, labels = eval_preds

examples/pytorch/multiple-choice/run_swag_no_trainer.py

@@ -31,10 +31,11 @@
 
 import datasets
 import torch
-from datasets import load_dataset, load_metric
+from datasets import load_dataset
 from torch.utils.data import DataLoader
 from tqdm.auto import tqdm
 
+import evaluate
 import transformers
 from accelerate import Accelerator
 from accelerate.logging import get_logger
@@ -514,7 +515,7 @@ def preprocess_function(examples):
             accelerator.init_trackers("swag_no_trainer", experiment_config)
 
     # Metrics
-    metric = load_metric("accuracy")
+    metric = evaluate.load("accuracy")
 
     # Train!
     total_batch_size = args.per_device_train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps

examples/pytorch/question-answering/run_qa.py

@@ -25,8 +25,9 @@
 from typing import Optional
 
 import datasets
-from datasets import load_dataset, load_metric
+from datasets import load_dataset
 
+import evaluate
 import transformers
 from trainer_qa import QuestionAnsweringTrainer
 from transformers import (
@@ -593,7 +594,7 @@ def post_processing_function(examples, features, predictions, stage="eval"):
         references = [{"id": ex["id"], "answers": ex[answer_column_name]} for ex in examples]
         return EvalPrediction(predictions=formatted_predictions, label_ids=references)
 
-    metric = load_metric("squad_v2" if data_args.version_2_with_negative else "squad")
+    metric = evaluate.load("squad_v2" if data_args.version_2_with_negative else "squad")
 
     def compute_metrics(p: EvalPrediction):
         return metric.compute(predictions=p.predictions, references=p.label_ids)

examples/pytorch/question-answering/run_qa_beam_search.py

@@ -25,8 +25,9 @@
 from typing import Optional
 
 import datasets
-from datasets import load_dataset, load_metric
+from datasets import load_dataset
 
+import evaluate
 import transformers
 from trainer_qa import QuestionAnsweringTrainer
 from transformers import (
@@ -625,7 +626,7 @@ def post_processing_function(examples, features, predictions, stage="eval"):
         references = [{"id": ex["id"], "answers": ex[answer_column_name]} for ex in examples]
         return EvalPrediction(predictions=formatted_predictions, label_ids=references)
 
-    metric = load_metric("squad_v2" if data_args.version_2_with_negative else "squad")
+    metric = evaluate.load("squad_v2" if data_args.version_2_with_negative else "squad")
 
     def compute_metrics(p: EvalPrediction):
         return metric.compute(predictions=p.predictions, references=p.label_ids)

examples/pytorch/question-answering/run_qa_beam_search_no_trainer.py

@@ -29,10 +29,11 @@
 import datasets
 import numpy as np
 import torch
-from datasets import load_dataset, load_metric
+from datasets import load_dataset
 from torch.utils.data import DataLoader
 from tqdm.auto import tqdm
 
+import evaluate
 import transformers
 from accelerate import Accelerator
 from accelerate.logging import get_logger
@@ -680,7 +681,7 @@ def post_processing_function(examples, features, predictions, stage="eval"):
         references = [{"id": ex["id"], "answers": ex[answer_column_name]} for ex in examples]
         return EvalPrediction(predictions=formatted_predictions, label_ids=references)
 
-    metric = load_metric("squad_v2" if args.version_2_with_negative else "squad")
+    metric = evaluate.load("squad_v2" if args.version_2_with_negative else "squad")
 
     def create_and_fill_np_array(start_or_end_logits, dataset, max_len):
         """

examples/pytorch/question-answering/run_qa_no_trainer.py

@@ -29,10 +29,11 @@
 import datasets
 import numpy as np
 import torch
-from datasets import load_dataset, load_metric
+from datasets import load_dataset
 from torch.utils.data import DataLoader
 from tqdm.auto import tqdm
 
+import evaluate
 import transformers
 from accelerate import Accelerator
 from accelerate.logging import get_logger
@@ -696,7 +697,7 @@ def post_processing_function(examples, features, predictions, stage="eval"):
         references = [{"id": ex["id"], "answers": ex[answer_column_name]} for ex in examples]
         return EvalPrediction(predictions=formatted_predictions, label_ids=references)
 
-    metric = load_metric("squad_v2" if args.version_2_with_negative else "squad")
+    metric = evaluate.load("squad_v2" if args.version_2_with_negative else "squad")
 
     # Create and fill numpy array of size len_of_validation_data * max_length_of_output_tensor
     def create_and_fill_np_array(start_or_end_logits, dataset, max_len):

examples/pytorch/question-answering/run_seq2seq_qa.py

@@ -25,8 +25,9 @@
 from typing import List, Optional, Tuple
 
 import datasets
-from datasets import load_dataset, load_metric
+from datasets import load_dataset
 
+import evaluate
 import transformers
 from trainer_seq2seq_qa import QuestionAnsweringSeq2SeqTrainer
 from transformers import (
@@ -581,7 +582,7 @@ def preprocess_validation_function(examples):
         pad_to_multiple_of=8 if training_args.fp16 else None,
     )
 
-    metric = load_metric("squad_v2" if data_args.version_2_with_negative else "squad")
+    metric = evaluate.load("squad_v2" if data_args.version_2_with_negative else "squad")
 
     def compute_metrics(p: EvalPrediction):
         return metric.compute(predictions=p.predictions, references=p.label_ids)

examples/pytorch/semantic-segmentation/run_semantic_segmentation.py

@@ -21,7 +21,6 @@
 from dataclasses import dataclass, field
 from typing import Optional
 
-import datasets
 import numpy as np
 import torch
 from datasets import load_dataset
@@ -30,6 +29,7 @@
 from torchvision import transforms
 from torchvision.transforms import functional
 
+import evaluate
 import transformers
 from huggingface_hub import hf_hub_download
 from transformers import (
@@ -337,7 +337,7 @@ def main():
     label2id = {v: str(k) for k, v in id2label.items()}
 
     # Load the mean IoU metric from the datasets package
-    metric = datasets.load_metric("mean_iou")
+    metric = evaluate.load("mean_iou")
 
     # Define our compute_metrics function. It takes an `EvalPrediction` object (a namedtuple with a
     # predictions and label_ids field) and has to return a dictionary string to float.

examples/pytorch/semantic-segmentation/run_semantic_segmentation_no_trainer.py

@@ -24,13 +24,14 @@
 import datasets
 import numpy as np
 import torch
-from datasets import load_dataset, load_metric
+from datasets import load_dataset
 from PIL import Image
 from torch.utils.data import DataLoader
 from torchvision import transforms
 from torchvision.transforms import functional
 from tqdm.auto import tqdm
 
+import evaluate
 import transformers
 from accelerate import Accelerator
 from accelerate.logging import get_logger
@@ -500,7 +501,7 @@ def preprocess_val(example_batch):
     args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
 
     # Instantiate metric
-    metric = load_metric("mean_iou")
+    metric = evaluate.load("mean_iou")
 
     # We need to initialize the trackers we use, and also store our configuration.
     # We initialize the trackers only on main process because `accelerator.log`

examples/pytorch/speech-recognition/run_speech_recognition_ctc.py

@@ -28,8 +28,9 @@
 import datasets
 import numpy as np
 import torch
-from datasets import DatasetDict, load_dataset, load_metric
+from datasets import DatasetDict, load_dataset
 
+import evaluate
 import transformers
 from transformers import (
     AutoConfig,
@@ -643,7 +644,7 @@ def is_audio_in_length_range(length):
     # instantiate a data collator and the trainer
 
     # Define evaluation metrics during training, *i.e.* word error rate, character error rate
-    eval_metrics = {metric: load_metric(metric) for metric in data_args.eval_metrics}
+    eval_metrics = {metric: evaluate.load(metric) for metric in data_args.eval_metrics}
 
     # for large datasets it is advised to run the preprocessing on a
     # single machine first with ``args.preprocessing_only`` since there will mostly likely

examples/pytorch/speech-recognition/run_speech_recognition_seq2seq.py

@@ -27,8 +27,9 @@
 
 import datasets
 import torch
-from datasets import DatasetDict, load_dataset, load_metric
+from datasets import DatasetDict, load_dataset
 
+import evaluate
 import transformers
 from transformers import (
     AutoConfig,
@@ -425,7 +426,7 @@ def is_audio_in_length_range(length):
         return
 
     # 8. Load Metric
-    metric = load_metric("wer")
+    metric = evaluate.load("wer")
 
     def compute_metrics(pred):
         pred_ids = pred.predictions

examples/pytorch/summarization/run_summarization.py

@@ -27,8 +27,9 @@
 import datasets
 import nltk  # Here to have a nice missing dependency error message early on
 import numpy as np
-from datasets import load_dataset, load_metric
+from datasets import load_dataset
 
+import evaluate
 import transformers
 from filelock import FileLock
 from transformers import (
@@ -598,7 +599,7 @@ def preprocess_function(examples):
     )
 
     # Metric
-    metric = load_metric("rouge")
+    metric = evaluate.load("rouge")
 
     def postprocess_text(preds, labels):
         preds = [pred.strip() for pred in preds]

examples/pytorch/summarization/run_summarization_no_trainer.py

@@ -30,10 +30,11 @@
 import nltk
 import numpy as np
 import torch
-from datasets import load_dataset, load_metric
+from datasets import load_dataset
 from torch.utils.data import DataLoader
 from tqdm.auto import tqdm
 
+import evaluate
 import transformers
 from accelerate import Accelerator
 from accelerate.logging import get_logger
@@ -583,7 +584,7 @@ def postprocess_text(preds, labels):
             accelerator.init_trackers("summarization_no_trainer", experiment_config)
 
     # Metric
-    metric = load_metric("rouge")
+    metric = evaluate.load("rouge")
 
     # Train!
     total_batch_size = args.per_device_train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps

examples/pytorch/text-classification/run_glue.py

@@ -25,8 +25,9 @@
 
 import datasets
 import numpy as np
-from datasets import load_dataset, load_metric
+from datasets import load_dataset
 
+import evaluate
 import transformers
 from transformers import (
     AutoConfig,
@@ -480,9 +481,9 @@ def preprocess_function(examples):
 
     # Get the metric function
     if data_args.task_name is not None:
-        metric = load_metric("glue", data_args.task_name)
+        metric = evaluate.load("glue", data_args.task_name)
     else:
-        metric = load_metric("accuracy")
+        metric = evaluate.load("accuracy")
 
     # You can define your custom compute_metrics function. It takes an `EvalPrediction` object (a namedtuple with a
     # predictions and label_ids field) and has to return a dictionary string to float.